How to measure AI share of voice

AI share of voice is the percentage of brand mentions in AI answers that belong to you, measured across the questions your buyers ask. If assistants name five brands in your category and one in four of those mentions is you, your share of voice is 25%.

It answers a different question from mention rate. Mention rate says how often you appear at all. Share of voice says how much of the conversation you hold against competitors. Our AI visibility metrics guide covers all five metrics side by side; this post is only about calculating the one most teams report to leadership.

The idea comes from advertising, where share of voice meant your share of all ad spend or impressions in a market. In AI search there are no ad slots to count. You count names inside generated answers instead.

The basic formula is simple:

> Share of voice = mentions of your brand ÷ mentions of all tracked brands, across the same set of answers

Count a mention once per answer. If ChatGPT names you three times in one long answer, that is still one answer that mentions you. Counting every repetition rewards verbose answers, not visibility.

Illustration: you track 40 prompts in one engine and run each prompt three times, giving 120 answers. Across those answers your brand is mentioned in 30, Competitor A in 48, Competitor B in 36 and Competitor C in 6. Total mentions: 120. Your share of voice is 30 ÷ 120 = 25%.

Two variants are worth knowing:

• Position-weighted share of voice. Give a mention more weight the earlier it appears, for example 1.0 for first, 0.7 for second, 0.5 for third and 0.3 after that. Being named first in a shortlist is worth more than being an afterthought. The original GEO research used a position-adjusted measure for the same reason (Aggarwal et al.).
• Prompt-weighted share of voice. Weight each prompt by how much it matters to revenue, for example 3 for a purchase prompt and 1 for a general information prompt. This stops 20 easy awareness questions from hiding a loss on the five prompts that close deals.

Pick one variant and keep it. Switching formulas between reports makes every trend meaningless.

Only prompts where a buyer could reasonably get a brand recommendation. Share of voice is only as good as the prompt set under it, so build that first.

Include:

• category prompts, such as "best invoicing software for freelancers";
• comparison prompts, such as "X vs Y for a 50-person team";
• alternative prompts, such as "alternatives to X";
• problem prompts where a product is the natural answer, such as "how do I automate supplier invoices?".

Leave out prompts that never name brands, such as definitions. They add answers with zero mentions and dilute nothing, but they also tell you nothing. Leave out branded prompts about you alone ("is [your brand] good?") from the main score. Track those separately for brand sentiment.

Thirty to fifty prompts per product line and market is enough to see a stable pattern. Prompt research helps you find which questions buyers actually ask.

Track the brands that actually appear in answers, not only the ones your sales team names. Your share of voice changes depending on who is in the denominator, so the competitor list must be honest and stable.

Start with a discovery run. Ask your prompts once in each engine and list every brand that appears. You will often find two kinds of surprise: an old competitor you no longer think about, and a marketplace, directory or review site that engines name as if it were a brand.

Then fix a list of five to eight brands for reporting. Add new entrants when they start appearing regularly, and note the date, because adding a brand lowers everyone's share overnight. A dip caused by a bigger denominator is not a loss of visibility.

Calculate one share of voice for each engine and each market, then decide whether a blended number helps. Usually a blended score hides more than it shows.

Engines behave differently. Perplexity searches the web for almost every answer, so it reacts quickly to new pages and reviews. ChatGPT mixes answers from model memory with answers from live search (OpenAI). Google's AI features draw on its search index (Google Search Central). A brand can hold 40% in one engine and 10% in another, and the average of 25% describes neither.

Markets differ even more. The same question asked in Turkish for Türkiye and in English for the US can produce a completely different shortlist, with local competitors you never see at home. Calculate each market from prompts written in its own language.

A simple reporting grid works well:

Illustrative values. Each cell is its own calculation. Leadership reads the grid in seconds and sees where to act.

More than one. AI answers vary between runs, so a single answer per prompt makes your share of voice jump around for no reason.

Three runs per prompt per measurement is a sensible minimum. Daily tracking smooths the noise further: a seven-day rolling share of voice is far steadier than any single day. Keep the run conditions the same each time: logged out, no chat history, the same location and language settings. Your own ChatGPT history, for example, can skew what you see.

Read it weekly, compare it with your own past, and look for the cause behind every move. The level matters less than the direction.

When the number changes, check these in order:

1. Did the denominator change? A new competitor in the list lowers everyone.
2. Did the engine change? Model updates can shift many answers at once.
3. Did the sources change? A review site update or a new listicle can add or remove you in many answers.
4. Did you change something? A new comparison page or a fixed pricing page can show up within weeks in engines that search live.

Keep the full answers on file so you can see which prompts moved. A 5-point drop that comes from three lost purchase prompts needs a different fix from a drop spread thinly across 30 information prompts.

Most errors come from inconsistency, not from maths. The common ones:

• Changing the prompt set every month. Add prompts in batches and report the old and new sets separately for one cycle.
• Counting your own branded prompts. They inflate the score and say nothing about competition.
• Mixing markets. A blended global number lets a strong home market hide a weak export market.
• One run per prompt. Variation looks like movement.
• Treating 100% as the goal. In most categories engines list several options. Being one of the first two names in the answers that matter is the realistic target.

Show one number per engine and market, the trend over the last 8 to 12 weeks, and the three prompts that moved most. Then name the action you are taking on each.

Share of voice is persuasive because it is relative. "We hold 25% of AI recommendations in our category; Competitor A holds 40%" is a sentence any executive understands. Pair it with the sources behind the gap, which citation tracking shows, and the plan writes itself. For a full work plan, see how to improve AI visibility in 30 days.

Divide your brand's mentions by the mentions of all tracked brands, counting once per answer, on a fixed set of buyer prompts. Fix the competitor list, run each prompt several times, calculate per engine and per market, and read the weekly trend. Ossvy's competitor analysis does the counting daily, or start with the free AI visibility check for a first snapshot.